Research On The Evaluation Theory And Method For Optical Flow Estimation Of Image Sequence

Optical flow estimation technique of image sequence is an important research field of computer vision. Since twenty-first Century, with the development of computer hardware and software level unceasing enhancement, there emerges an endless stream of new theory and new method for optical flow computation, which continue to expand the scope of the optical flow calculation theory in the practical production and the application of life, such as the application in robot vision, aerospace, medical image processing, weather analysis, unmanned vehicle autonomous driving and other fields.With the rapid development of optical flow computation technology, it has become an important content of research that how to measure the performance of optical flow algorithm on accuracy and robustness and so on, to establish perfect optical flow evaluation system, and to carry on objective and comprehensive evaluation of different optical flow calculation theories and methods. Although current evaluation theories and methods provide affluent image sets and evaluation indexes that can offer good performance evaluation system for the calculation of optical flow, there are still some limitations in the existing optical flow error evaluating theories, which cannot fully and accurately reflect the performance differences. In view of the above problems, this paper put forward a new evaluation theory based on the optical flow vector, so it can more comprehensively and objectively reflect the error characteristics and performance of optical flow algorithms. This paper mainly includes the following contents:1 aiming at the limitation of the angle error is analyzed, and the calculation of angle error is set from 3D to 2D. While optical flow vector is 2D data, the 3rd coordinate introduced by researchers is lack of rationality, and it also causes underestimation of real angle error and evaluation change caused by optical flow magnitude etc.. This paper, therefore, redefines the angle error and directly calculate 2D angle error between the evaluation of optical flow and the real flow.2 In the existing evaluation methods, the endpoint error is the modulus value of vector difference between evaluated flow and the real. That’s mean it includes the measure of angle error, and cannot accurately reflect the magnitude difference of them. On this account, this paper define the magnitude difference as endpoint error, avoiding the angle error’s affect. Also, this paper proposes normalized endpoint error, which takes the modulus value of real flow into account, and helps to objectively reflect the real flow relative to the difference error.3 Based on these new error definitions, this paper improves the statistical methods of error in the theory of optical flow assessment. The expected value rather than the average is used to represent the overall level of optical flow error. Moreover, robustness parameters are also redefined to adjust new error evaluation method.4 With the standard image sequences for optical flow evaluation in Middlebury, this paper tests this new evaluation theory, and compare it with existing evaluation method, to prove its objectivity and comprehensiveness.